OpenAI reports that updated safeguards in GPT-6.2 Astra and GPT-6.2 Sol reduced the average attack success rate for protected reasoning extraction from 54.0% to 0.2% during internal adaptive red-teaming. The company released these findings alongside a broader breakdown of its defense stack against adversarial distillation, a technique where unauthorized actors train secondary models on a frontier model's outputs to bypass safety controls.
Adversarial distillation poses a direct safety risk in dual-use domains like cybersecurity and biology. A model trained to help defenders find software vulnerabilities can learn to exploit them if the underlying capabilities transfer without the original model's usage restrictions. OpenAI's internal study on capability transfer used trajectories from Astra-level teacher models working on coding and defensive cybersecurity tasks. Each trajectory included tool interactions, final responses, and exact reasoning traces. Student models fine-tuned on these exact traces consistently outperformed students trained on reasoning summaries after a lightweight round of reinforcement learning, particularly at larger inference budgets. The results indicate that protecting reasoning traces limits the transfer of risky cyber capabilities, even when the base tasks fall under standard usage policies.
OpenAI defends against these extraction attempts across two complementary layers. At the request level, the system uses input classifiers to detect reasoning extraction prompts and output classifiers to catch leaks in assistant messages and tool-call arguments. These classifiers apply to both streaming and non-streaming responses. The API enforces strict boundaries on how protected reasoning can be reused. GPT-6.2 Astra and Sol now tie reasoning traces directly to the conversation history that produced them. If a user rewrites an earlier message, the API drops the reasoning that followed it. The system also prevents older model families from reading the protected reasoning of newer models, closing a pathway that previously allowed encrypted reasoning to leak across user and workspace boundaries.
At the campaign level, OpenAI monitors request-level and account-level signals to identify coordinated extraction attempts. The company enforces against networks of accounts violating its terms of service, with weekly enforcement actions ranging from thousands to tens of thousands of accounts. OpenAI attributes core components of recent distillation campaigns to individuals associated with DeepSeek and Moonshot AI, the developer of Kimi. The company shares indicators of compromise with other model providers and supports legislation like the Chip Security Act and the AI OVERWATCH Act to restrict foreign actors from illicitly extracting capabilities from U.S. frontier models.
LLM-assisted peer review collapses critique diversity
An unreviewed arXiv paper published in December 2025 documents a measurable decline in the independence of academic peer reviews, attributing the shift to widespread LLM use. The study, titled "Twice-Told Critiques: Claim-Level Homogenization of Peer Review in the LLM Era," measures how often reviewers of the same paper raise identical problems. The authors report that critique overlap at ICLR rose from 12.0% in 2026 to 21.3% in 2027. Similar year-on-year increases appeared at NeurIPS, ICML, and several major journals.
Illustrative image: geralt / Pixabay
The researchers built a pipeline to extract and group critique claims from review texts. They used GPT-6 Luna to parse individual reviews into discrete problems, excluding praise and neutral text. A separate grouping step then clustered claims across a panel of reviews for a single paper. To identify synthetic submissions, they applied MELD, a detector trained to score reviews for LLM-generated or edited text. The detector flags reviews above the 99.9th percentile of pre-2023 scores. Within a single paper, pairs of reviews flagged by MELD shared significantly more problems than pairs of unflagged reviews. At ICLR, the overlap gap between flagged and unflagged pairs widened from 3.0 points in 2024–2025 to 4.5 points in 2026–2027.
The homogenization effect extends beyond wording. When the authors measured distinct claims per three-review panel at ICLR, the average number of unique problems fell from 19.9 in 2026 to 19.0 in 2027. A third reviewer added 5.46 new problems to a panel in 2027, down from 6.12 in 2026. The study also tracked topic shifts, noting that clarity critiques dropped from 24% of claims in 2020–2022 to 10% in 2027, while requests regarding experiment scope and statistical rigor rose.
The authors acknowledge several limitations in their methodology. MELD flags 39.5% of human reviews that were polished by LLMs, meaning a flag does not strictly separate generation from editing. The detector's threshold may also be too low if its training data included the authors' own pre-2023 reviews. Furthermore, the rise in overlap was not confined to flagged reviews. Overlap among unflagged pairs increased from 10.8% to 14.3% over the same period, suggesting that factors beyond detectable LLM text generation contribute to the convergence of critiques. The authors note that a reviewer could extract critique points from an LLM and rewrite them in their own voice, a pattern that text detectors cannot capture.
The paper concludes that the independence of expert critiques is degrading in both content and style. While LLM-flagged submissions drive a substantial portion of the measured overlap, the broader trend points to a systemic shift in how academic feedback is generated. The authors recommend that venues continue monitoring critique diversity as LLM assistance becomes standard practice in the review process.
Sources
- openai.com: Combatting adversarial distillation
- arXiv: Twice-Told Critiques: Claim-Level Homogenization of Peer Review in the LLM Era
PNEUMETRON EDITORIAL
Pneumetron's pipeline pairs AI-assisted drafting with human editorial review before publishing — our goal is to make staying informed easier, not to replace real reporting.




